Speech Coding Standards

Antti KiviluotoAntti.Kiviluoto@hut.fi

Speech coding standards –methods thatenable voice communication

Presentation• Intro• Speech Analysis - Synthesis & Linear Prediction• Speech coding standards• Subband & Transform Coders• Applications for synthetic speech• Summary

Intro• Interest towards speech coding & standardization:

– World wide growth in communication networks– Emergence of new multimedia applications– Advances in Very Large-Scale Integration (VLSI)

devices• Standardization

– International Telecommunications Union (ITU)– European Telecom. Standards Institute (ETSI)– International Standards Organization (ISO)– Telecommunication Industry Association (TIA), NA– R&D Center for Radio systems (RCR), Japan

Intro• Signal models are based on Linear Predictive

Coding (LPC)– Open-loop linear prediction (LP)– Closed-loop/Analysis-by-synthesis LP– Subband & transform coding algorithms

• Quality standards for SCA– Mean Opinion Score (MOS)– Diagnostic Rhyme Test (DRT)– Diagnostic Acceptability Measure (DAM)

Speech Analysis - Synthesis & Linear Prediction

• Speech is produced by the interaction of thevocal tract with the vocal chords in the glottis

• Engineering model:– Vocal tract - time-varying digital filter– Filter parameters estimated by using LP algorithms

Speech Analysis - Synthesis & Linear Prediction

• Linear Prediction– Introduced in late 1960’s– Widely used in speech signal processing– Number of available techniques since introduction– Basic principle of prediction still the same– Most recent sample predicted by a linear

combination of past samples– LP analysis done by Finite-length Impulse

Response (FIR) digital filter

Speech Analysis - Synthesis & Linear Prediction

• Linear prediction analysis

Speech Analysis - Synthesis & Linear Prediction

• Open-loop linear prediction (waveform)– Low complexity codecs– Good quality with rates above 16 Kbit/s– Pulse Code Modulation (PCM)

• simplest form of waveform codecs• Non-linear quantization• quantizing the speech signal directly

– Adaptive Differential PCM (ADPCM)• quantize the difference between the speech signal and a

prediction

Speech Analysis - Synthesis & Linear Prediction

• Closed-loop/Analysis-by-synthesis LP– introduced in 1982– Intelligible speech at 2.4 kbits/s and below– Frame size usually 20 ms– Cannot provide “natural” sounding speech at any bit

rate– Multi-Pulse Excitation (MPE) codec– Regular-Pulse Excitation (RPE) codec– Code-Excitation Linear Predictive (CELP) codec– Vector Sum Excited Linear Prediction (VSELP)

Speech Analysis - Synthesis & Linear Prediction• Splitting the input speech to

be coded into frames• Parameters determined for

Synthesis filter• Excitation determined to the

filter• Encoder transmits

informationrepresenting the filterparametersand the excitation to thedecoder

• at the decoder the givenexcitation is passed throughthe synthesis filter to givethe reconstructed speech

• AbS Codec Structure

Speech Analysis - Synthesis & Linear Prediction

• Short-Term Prediction (STP)– Captures the formant structure of short-term

speech spectrum

• Long-Term Prediction (LTP)– Captures the long-term correlation in the speech

signal– Representing the periodicity in speech

Speech Analysis - Synthesis & Linear Prediction

• Multi-Pulse Excitation (MPE) codec– Skyphone standard of British Telecom International,

used in passenger communications in aircraft– Excitation is a sequence of multiple nonuniformly

spaced pulses– Uses both short- and long-term prediction

• Regular-Pulse Excitation (RPE) codec– Used in GSM standard– Excitation is a sequence of multiple uniformly

spaced pulses– Uses only long-term prediction

Speech Analysis - Synthesis & Linear Prediction

• Code- Excitation Linear Predictive (CELP) codec– Codebook contains excitation vector– Uses long- and short-term prediction

Speech Analysis - Synthesis & Linear Prediction

• Quality standards for SCA– Evaluation of speech quality and intelligibility

• Mean Opinion Score (MOS)– Quite common in standardization tests– Five-level quality scale, 5 (excellent) – 1 (bad)

• 4.0-4.5: network or toll quality• 3.5-4.0: communications quality (cellular grade)• 2.5-3.5: synthetic quality

Speech coding standards

• ITU G.726 and G.727 ADPCM Coders– Quantize the difference between the speech

(current) signal and a predicted speech samples– Prediction parameters are obtained by backward

estimation, from already quantized data– Algorithm operates at 16, 24, 32, 40 and 64 Kbit/s– G.726: individually optimized quantizers– G.727: embedded quantizers - developed for

packet network applications (switch rates)– MOS for 32 Kbit/s: 4.1– Complexity: 2 MIPS

Speech coding standards

• Inmarsat-B Adaptive Predictive Vocoder– The International Mobile Satellite B standard– Uses ADPCM coder with LTP and 6th order STP– Algorithm operates at 9.6 and 12.8 Kbit/s– MOS for 9.6 (12.8) Kbit/s: 3.1 (3.4)– Complexity: 10 MIPS

Speech coding standards

• LPC-10(e) Federal Standard 1015– Introduced by Department of Defense in 1976– Secure communications– Uses 10th order predictor– Algorithm operates at 2.4 Kbit/s– MOS for 2.4 Kbit/s: 2.3– Complexity: 7 MIPS

Speech coding standards

• Mixed Excitation Linear Prediction (MELP) codec– Introduced by U.S. government in 1996– Development motivated by voicing errors in LPC-10– Algorithmic delay estimated 122.5 ms– Algorithm operates at 2.4 Kbit/s– MOS for 2.4 Kbit/s: 3.2– Complexity: 40 MIPS

Speech coding standards

• ETSI full-Rate GSM 6.10– Pan-European digital cellular standard– Uses RPE coder with LTP– Algorithm operates at 13.0 Kbit/s– MOS for 13.0 Kbit/s: 3.5-3.9– Complexity: 5-6 MIPS

Speech coding standards

• FS-1016– Used by Department of Defense (DoD) for 3rd-

generation secure telephone unit (STU-III)– Developed jointly by DoD and AT&T Bell Laboratories– Uses CELP algorithm– excitation is formed by combining vectors from LTP

and stochastic codebook– Algorithm operates at 4.8 Kbit/s– MOS for 4.8 Kbit/s: 3.2– Complexity: 16 MIPS

Speech coding standards

• Vector Sum Excited LP (VSELP) Standards– Algorithm embedded to 3 digital cellular standards

• TIA IS-54 (Japan & North America)• ETSI 6.20 GSM half-rate (Europe)

– Highly structured codebooks• Reduceses computational complexity• Increases robustness to channel errors

– Algorithm operates at 2.4 Kbit/s– MOS for 6.3 (8.0) Kbit/s VSELP: 3.4 (3.5)– Complexity: 14 MIPS

Speech coding standards

• ITU G.728 Low-Delay CELP– Uses low-delay CELP coder

• Short frames• Short exication vectors• No LTP and the order of STP is increased to 50

– Speech quality equivalent or better than G.726– Algorithm operates at 16.0 Kbit/s– MOS for 16.0 Kbit/s: 3.4– Complexity: 30 MIPS

Speech coding standards

• IS-96– Used in Code Division Multiple Access (CDMA),

standard for cellular communications– Uses CELP algorithm with STP– Bit rate is variable, determined by speech activity– Lower rates achieved by allocating fewer bits to LP

parameters– Algorithm operates at 1.2, 2.4, 4.8 and 9.6 Kbit/s– MOS for 9.6 Kbit/s: 3.3– Complexity: 15 MIPS

Speech coding standards• ITU G.729 and G.729A CS-ACELP

– Designed for wireless and multimedia networkapplications

– Uses Conjugate Structure Algebraic CELP• Low-delay algorithm (15 ms)• Frame size 10 ms• Uses 2 codebooks

– G.729 Annex has lower complexity and quality– Algorithms are interoperable– Algorithm operates at 8.0 Kbit/s– MOS for 8.0 Kbit/s: G.729=4, G.729A=3.8– Complexity: G.729=20 MIPS, G.729A=11 MIPS

Speech coding standards

• ITU G.723.1 MP-MLQ/ACELP– Speech coder for audio and videoconferencing over

public telephone networks– Part of ITU H.323 and H.324 standards– Also recommended to be default audio codec for

voice of the network (VOIP)– Standard is dual-rate with 2 excitation schemes

• Multipulse maximum likelihood quantization for 6.3 Kbit/s• ACELP for 5.3 Kbit/s

– MOS for 5.3 (6.3) Kbit/s: 3.7 (4.0)– Complexity: 16 MIPS

Subband & Transform Coders

• Subband Coder– The signal is divided into frequency subbands– Coder exploits the statistics of the signal and encodes

each band using a different number of bits– E.G lower-frequency bands are allotted with more bits

• Preserve critical pitch and formant information

– ITU G.722, standard for ISDN teleconferencing• Low frequency subband quantized with 48 Kbit/s• High-frequency with 16 Kbit/s

– MOS for 64 Kbit/s: 4.1– Complexity: 5 MIPS

Subband & Transform Coders

• Sinusoidal Transform Coding– Speech is represented by a linear combination of

sinusoids– Opportunity to reduce the bit rate

• Speech is typically periodic• Lots of pauses

– Considered for wideband and high-fidelity app.– MOS for 2.4 (4.8) Kbit/s: 2.9 (3.5)– Complexity: 13 MIPS

Applications for synthetic speech• Telecommunications and Multimedia• For the Blind• For the Deafened and Vocally Handicapped• Educational• Samples…

Summary

20103.1/3.49.6/12.8Inmarsat-B APC

22.5403.22.4MELP

20143.58IS-54 VSELP

30163.24.8FS 1016 – CELP

20153.31.2/2.4/4.8/9.6IS-96 QCELP

20143.46.3GSM HR VSELP

10114.0/3.76.3/5.3MPC-MLQ G.723.1

1020 (11)4.0 (3.8)8CS-ACELP(-A) G.729

0.625304.016LD-CELP G.728

0.12554.148/56/64SBC G.722

0.12524.132ADPCM G.726

00.014.364PCM G.711

Frame Size(ms)

Complexity(MIPS)

MOSBit Rate(Kbit/s)

Algorithm

Summary• Linear prediction widely used in speech coding• Speech coding research has come a long way in

the last 10 years• Need for speech codecs from cellular phones to

multimedia systems

The End

• Time for questions